Dual-Stage Domain Adaptation with Semantic Center Alignment for Cross-Subject EEG-Based Depression Screening

Current diagnosis of Major Depressive Disorder (MDD) relies heavily on subjective clinical interviews, lacking objective physiological markers. While EEG-based deep learning holds promise, its clinical deployment is severely hindered by inter-subject variability, where models fail to generalize to unseen patients. To bridge this gap, we propose a dual-stage domain adaptation framework combining ResNet-BiLSTM with Domain Adversarial Neural Networks (DANN) and Semantic Center Alignment (SCA) for robust cross-subject depression detection. We transform 128-channel EEG into topology-preserving maps to retain pathologically relevant spatial connectivity and employ a hybrid backbone to capture spatiotemporal features. Specifically, we integrate a two-level adaptation strategy: DANN aligns global feature distributions across subjects via adversarial training, while SCA enforces class-specific compactness by maintaining dynamic prototype centers in a memory bank, mitigating negative transfer. Extensive experiments on the MODMA dataset demonstrate that our method significantly outperforms subject-dependent baselines (59.18%), achieving 79.09% accuracy with improved stability across diverse subjects. By effectively addressing domain shifts through this dual mechanism, our framework offers a more generalizable screening tool that substantially reduces the reliance on individual calibration, paving the way for reliable computer-aided diagnosis.